Every serious research question spans corpora that no single product bundles together.
A materials team wants computed properties next to the mines that produce the input and the export designations on the suppliers. A market-access consultancy wants procedure payment rates next to device clearances and enforcement history. A power developer wants queue positions next to generating units next to the rule that changes the timeline. Each of those is three or four datasets, published by three or four different agencies, in three or four different shapes.
The usual options are both bad. Pour a pile of PDFs into a general-purpose assistant and you get fluent answers that cannot be traced to a filing. Or commission a retrieval integration per source and spend a quarter on plumbing before anyone answers a question.
Today we are shipping a third option: build the model yourself, from the catalogue, in one sitting.
Five steps, and only one of them costs anything
| Step | What you do | What it costs |
|---|---|---|
| 1 · Pick | Choose datasets from the catalogue — filter by area, or search by name. Each card names the publishing body, what one record is, and who it is for. | Free |
| 2 · Group | Put up to three datasets in a group, and up to four groups in a model. A group is one line of enquiry: its datasets are searched together. | Free |
| 3 · Tune | Set how many records each dataset returns. Optionally rename the model, describe what it is for, and edit how any dataset is written up. | Free |
| 4 · Make it live | The model gets its own code. It is live immediately — no review queue — and reachable only by you until you publish a channel. | Free to build |
| 5 · Call it | From the workspace, the API, or an embedded widget on your own site. | Per call — see below |
The catalogue: 41 datasets, each with a stated grain
Every card tells you three things before you spend anything — who publishes it, what one record is, and who tends to need it. The second matters more than it looks: the record grain is exactly what the “records returned” dial counts, so it is the difference between asking for ten papers and ten passages of one rule.
| Area | Datasets | Examples of what one record is |
|---|---|---|
| Energy & grid | 12 | One project in an interconnection queue · one generating unit · one wholesale power contract · one rate plan · one power plant · one offshore wind lease |
| Regulatory & trade | 10 | One device clearance decision · one enforcement action against a named firm · one restricted party with its source list · one passage of a rule or guidance · one excluded provider |
| Life sciences | 6 | One published paper · one preprint · one measured binding affinity for a ligand–target pair · one chemical entity · one disease concept · one approved drug with its indications |
| Innovation & funding | 5 | One federal award action · one SBIR/STTR award · one certified labor condition application · one robotics paper · one notable trained model |
| Materials & minerals | 5 | One commodity–company–site mineral operation · one mine · one commodity×country×year statistic · one computed material · one county water-use profile |
| Healthcare economics | 3 | One drug×plan-year spending row · one procedure code with its payment rate · one provider×procedure volume row |
These are authoritative sources, not scrapes: the national library’s biomedical literature, the medicines regulator’s clearance decisions, the energy agency’s generator inventory, the geological survey’s commodity statistics, the trade administration’s restricted-party lists, the treasury’s award records. Some are large — over two million measured binding affinities, over two and a half million wholesale power contract records, 1.3 million certified labor filings. Some are small and precise — 26,660 disease concepts, 96 nuclear generating units, 52 offshore wind leases. Both kinds earn their place; a narrow registry that answers one question exactly is worth more than a big index that approximates.
Groups are how you shape the thinking
A group is one line of enquiry. Its datasets are searched at the same time and read together, so datasets that answer the same question belong in the same group. Groups run one after another, each adding roughly fifteen seconds — and each adding to the price.
| Quality tier the call runs at | Credits per filled group | A 3-group model costs |
|---|---|---|
| Fast | 5 | 15 credits per call |
| Standard | 10 | 30 credits per call |
| Premium | 30 | 90 credits per call |
Three things are worth saying plainly about that table. Building is free and stays free no matter how long you spend tuning. Adding datasets inside a group is free, and so is raising how many records each one returns — only adding a group changes the price. And the tier is chosen per call, not per model, so the same model costs 5, 10 or 30 credits a group depending on how it is invoked. The builder quotes all three so nobody is quietly billed at the most expensive one.
Every dataset arrives with its own guardrail
This is the part that is hard to build yourself, and the part that decides whether the output survives contact with a reviewer.
Each dataset ships with two pieces of written guidance. The instruction says how its records should be presented and cited — mechanical, and easy. The guardrail says what that dataset does not support, and that is domain knowledge rather than boilerplate:
- A preprint has not been peer reviewed, may be revised or withdrawn, and must never outweigh peer-reviewed evidence in the same answer.
- A chemical ontology defines what a substance is. It carries no dosing, effect or hazard information, and none may be inferred from an entry.
- A device clearance is a substantial-equivalence finding, not a safety endorsement.
- An interconnection queue position is a request, not a plant.
- Retrieval is not appraisal: a paper matching a query is not evidence that its conclusion is correct, replicated or current.
You can override either half on any dataset in your model, and your text is stored with your model — so a later change to the shared template cannot silently rewrite prose you tuned.
There is also an off-topic gate, which declines questions outside the subject area your model covers. You can name the subject area, switch the gate off if your datasets answer more than one thing, and set who the refusal speaks as — because on a white-label deployment the one message a declined user actually reads should carry your name, not ours.
Live immediately. Private by default. Yours to retire.
- No review queue. A model is validated as you build it, so making it live is instant.
- Private until you say otherwise. A live model is reachable only by its owner. Publishing is a deliberate act, per channel — the API, an embedded widget, or both.
- The code is stable. Editing a live model recomposes it in place; the code your integrations point at does not change, so tuning a model never forks its identity.
- Retire and restore. Retiring a model keeps the row and the code, and callers get an unbilled refusal. Bring it back whenever you want. Deleting is permanent, asks twice, and tells you exactly what will happen first.
- Citations resolve. Answers cite the record, not the dataset in general — the filing, the paper, the award, the lease.
Try it before you have an account
The builder is open at /labs/model-builder/new, signed out. You assemble a real model against the real catalogue and see the real price. Saving and making it live need an account — and the draft you built is carried through signup, so nothing is lost.
We built it that way for one reason: the only question worth asking of a model builder is whether your sources are in it, and a mock with invented data answers a different question dishonestly.
Who is this for
- Consultancies. Your framework is the value; gathering the evidence is the cost. Assemble the sources your practice actually depends on once, then run every engagement through them — with sources your client can open.
- Product teams. Put a domain-accurate answer box in your own product, on your own datasets-of-record, without staffing a retrieval team.
- Diligence and corporate development. One model that spans the registry, the literature and the funding record beats three tabs and a spreadsheet.
- White-label partners. Your branding, your subject area, your refusal message.
What it does not do yet
- You cannot upload your own corpus to it today. A model is assembled from the published catalogue. Bringing a proprietary dataset in is a data-onboarding conversation, not a self-serve button — we would rather say so than let you discover it after you have built half a model.
- The catalogue is the catalogue. It grows when a new dataset is onboarded, and new datasets appear in the picker the moment they land — but you are choosing from what is there.
- Grouping is bounded on purpose: up to three datasets per group, up to four groups. Past that, models get slower and vaguer rather than better.
- A model is only as good as its picks. Nothing here rescues a model built from datasets that cannot answer the question. The card copy exists so you can tell before you build.
Open the builder at /labs/model-builder/new, or read the developer documentation for how to call one from your own product.